Using artificial intelligence models to generate dietary recommendations for chronic kidney disease patients: A comparative cross-sectional study.

BACKGROUND & AIMS: Artificial intelligence (AI) models are increasingly being used to assist in chronic kidney disease (CKD) dietary guidance, but concerns about their accuracy remain. This study aimed to assess the performance of three AI models (ChatGPT, DeepSeek, Gemini) in accordance with the Kidney Disease: Improving Global Outcomes (KDIGO) 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease.

METHODS

We compared ChatGPT, DeepSeek, and Gemini with the KDIGO 2024 guidelines through expert and cross-AI assessments. Outputs were scored across five domains: accuracy, clarity, completeness, practicality, and safety.

RESULTS

The AI models were able to generate well-structured dietary advice for CKD patients. DeepSeek received a higher score (7.625) for practicality compared to ChatGPT (6.5) and Gemini (7.0) (p = 0.041). ChatGPT showed numerically higher accuracy scores than the other two models, although there was no statistically significant difference (ChatGPT: 7.25, DeepSeek: 6.875, Gemini: 6.875; p = 0.715). Similarly, scores for clarity (p = 0.489), completeness (p = 0.432) and safety (p = 0.413) were not significantly different. The AI models share several common limitations: they might overlook contraindications (such as recommending a low-protein diet for patients with severe hypoalbuminemia), neglect age-specific needs, and lack real-time knowledge updates.

CONCLUSIONS

While AI shows clinical potential, its outputs still require expert validation, real-time updates, and enhanced safety protocols to ensure reliable clinical integration.